supervised classification applied to vegetation mapping in the barˆo de … · 2014. 5. 8. ·...

GEOGRAFIA, Rio Claro, v. 38, Número Especial, p. 9-23, ago. 2013.

SUPERVISED CLASSIFICATION APPLIED TO VEGETATIONMAPPING IN THE BARÃO DE MELGAÇO MUNICIPALITY (MATO

GROSSO STATE, BRAZIL), USING MODIS IMAGERY

Rafael Carlos BISPO1

Maria Angélica PETRINI2

Rubens Augusto Camargo LAMPARELLI3

Jansle Vieira ROCHA4

Abstract

The use of remote sensing images and geoprocessing techniques has contributed to afast and effective monitoring of the Pantanal region. Barão de Melgaço municipality, in Mato GrossoState, has 99.2% of its land area inserted in this biome, with flooding dynamics which result inrapid changes in vegetation cover. Accordingly, this study aimed to map the vegetation of thismunicipality, in both dry and rainy seasons, using supervised classification (Parallelepiped andSVM) of images from MODIS sensor. The classifications were applied, for each date, both oncomposite R(MIR) G(NIR) B(RED) and on composite with the fraction images R(wet soil) G(vegetation)B(water) for the rainy season, and R(dry soil) G(vegetation) B(water) for the dry season. Thefraction images were derived from a Linear Spectral Mixture Model. A total of eight maps werecreated and evaluated using an error matrix. The best result for each season was the Parallelepipedclassifier applied to the composition with the original bands from MODIS. These results werecompared with the map of phyto-ecological regions of Barão de Melgaço. They allowed concludingthat the presence of soil in Savanna Woodland and in Tree and/or Shrub Savanna prevented theclassification of these vegetation types in the images of the dry season.

Key-words: Digital classification. Error matrix. Linear spectral mixture model. Fractionimages.

1 Faculdade de Engenharia Agrícola - Feagri/Unicamp, Mestrando, Caixa Postal 6011 - 13083-875 - Campinas- SP, Brasil, E-mail: [email protected]

2 Faculdade de Engenharia Agrícola - Feagri/Unicamp, Doutoranda, Caixa Postal 6011 - 13083-875 - Campinas- SP, Brasil, E-mail: [email protected]

3 Núcleo Interdisciplinar de Planejamento Energético - Nipe/Unicamp, Pesquisador, 13083-860 - Campinas- SP, Brasil, E-mail: [email protected]

4 Faculdade de Engenharia Agrícola - Feagri/Unicamp, Professor Doutor, Caixa Postal 6011 - 13083-875 -Campinas - SP, Brasil, E-mail: [email protected]

10 GEOGRAFIASupervised classification applied to vegetation mapping in the Barão deMelgaço municipality (Mato Grosso state, Brazil), using MODIS imagery

Resumo

Classificação supervisionada aplicada ao mapeamento da vegetação do municípiode Barão de Melgaço/MT utilizando imagens do SENSOR MODIS

O uso de imagens de sensoriamento remoto e técnicas de geoprocessamento temcontribuído para o monitoramento rápido e eficaz do Pantanal. O município de Barão de Melgaço/MT tem 99,2% de sua área territorial inserida no bioma, com a dinâmica de inundação ocasionandomudanças rápidas na cobertura vegetal. Nesse sentido, o objetivo deste estudo foi mapear avegetação do município mato-grossense de Barão de Melgaço, nas estações seca e chuvosa,através de classificação supervisionada (Paralelepípedo e SVM) de imagens do sensor MODIS. Asclassificações foram aplicadas, para cada data, tanto na composição R(MIR) G(NIR) B(RED) quantona composição com as imagens-fração R(solo úmido) G(vegetação) B(água) para a estaçãochuvosa, e imagens-fração R(solo seco) G(vegetação) B(água) para a estação seca. As imagens-fração foram derivadas do Modelo Linear de Mistura Espectral. Ao todo foram gerados oitomapeamentos, avaliados por meio da matriz de erros. O melhor resultado para cada estação foio uso do classificador Paralelepípedo aplicado sobre a composição com as bandas originais doMODIS. Estas classificações foram comparadas com a carta de regiões fitoecológicas de Barão deMelgaço, permitindo avaliar que a presença de solo na Savana Arborizada e na Savana Gramíneo-Lenhosa influenciou para a não classificação das mesmas na estação seca.

Palavras-chave: Classificação digital. Matriz de erros. Modelo linear de mistura espectral.Imagens-fração.

INTRODUCTION

Pantanal is considered the largest floodable floodplain in the world, with a size ofapproximately 160,000 km2 and it is located in 3 countries, namely Brazil, Paraguay andBolivia (JUNK, 2006). According to Silva & Abdon (1998), the Brazilian Pantanal is inserted inthe Upper Paraguay river basin, encompassing an area of 138,183 km2 distributed in twoBrazilian States: 89,318 km2 in Mato Grosso do Sul, and 48,865 km² in Mato Grosso. As forBarão de Melgaço municipality, 99.2% of the territory is located within Pantanal and has theNatural Heritage Private Reserve SESC Pantanal, besides being part of the State Park Encontrodas Águas.

The flooding in Pantanal occurs typically between November and February and thedrought period extends from July to August. Adamoli (1995) reports that this flooding frequencyis a fundamental ecological factor of this biome, which determines the pulses of the mainbiotic and non-biotic processes, as well as operative cycles of productive processes, suchas cattle raising, tourism, and river navigation. Besides that, the flood regime causes thetype and the specific compositions of different landscape units occurring in the biome.

This particularity makes the monitoring of Pantanal a difficult task because, besidesits extension and access difficulties, this region suffers cyclic changes caused by changesof vegetation cover. Shimabukuro et al. (1998) inform that the monitoring of natural resourcesin regions like Pantanal can be improved using remote sensing methods.

Satellite image classification is one of the most frequent remote sensing applicationsfor the production of thematic maps, such as those of land use/land cover. This procedurecan be accomplished by both visual analysis and by automatically using a computer (FOODY,2002).

Among the automatic classifications, Devi & Baboo (2011) report that the so-calledParallelepiped algorithm is computationally efficient to classify data obtained from remote

11Bispo, R. C. / Petrini, M. A. /

Lamparelli, R. A. C. / Rocha, J. V.v. 38, Número Especial ago. 2013

sensors and that it is generally used when the process must be fast. Its disadvantage,however, in many cases, is its low accuracy.

On the other hand, the SVM (Support Vectors Machine) method is gaining prominencedue to its good performance related to other classifiers, as it was verified by Shao & Lunetta(2012) for the classification of land cover from the Albemarle-Pamlico estuarine system(USA), using temporal MODIS datasets and by Senthilnath et al. (2012), who combinedimage segmentation and classification to map the Krishna river, in southern India, and toevaluate its flooded regions, also using MODIS images.

The MODIS (Moderate Resolution Imaging Spectro-radiometer) sensor was designedto deliver simultaneous observations of atmospheric, oceanographic, and terrestrial features,at the visible and infrared wavelengths of the electromagnetic spectrum (JENSEN, 2009).Data from this sensor are available free of charge and its spectral bands, besides beinggeometrically, atmospherically, and radiometrically corrected, are also transformed into otherproducts such as vegetation indices and leaf area indices (LATORRE et al., 2007).

Zhan et al. (2002) inform that the red and infrared wavelength intervals are the mostimportant spectral regions to map vegetation. These bands are found in the module MOD13Q1from MODIS.

MODIS images and its products have been used in studies at Pantanal, e.g. monitoringof its spatial-temporal dynamics (ADAMI et al., 2008); to evaluate the performance of NDVI(Normalized Difference Vegetation Index), EVI (Enhanced Vegetation Index) and LSWI (LandSurface Vegetation Index), considering the main types of land cover and the seasonality ofrainfall in Pantanal, the LSWI was the most adequate index for this region (VIANA; ALVALÁ,2011); and also for the analysis of the vegetation phenology based on a temporal series ofEVI (ANTUNES et al., 2011).

Although MODIS data have a high temporal resolution, its moderate spatial resolutionfrequently becomes a limitation for the automatic image classification, because the spectralresponse of each pixel is the result of radiance integration from several targets, causing thepredominance of an intra-pixel spectral mixture (SILVA et al., 2010). According to Shimabukuro& Smith (1991), it is possible to decompose this spectral response using the Linear SpectralMixture Model (LSMM), which generates fraction images representing the proportions ofeach component in the pixel formation.

In order to monitor the dynamics of vegetation cover in the Pantanal region,Shimabukuro et al. (1998) used, based on Landsat-TM data, vegetation index images andfraction images derived from a LSMM, indicating a higher sensitivity of the fraction image tovariations of the vegetation cover when compared with the vegetation index image.

In this context, the use of fraction images for the automatic image classification,especially those with low to moderate spatial resolution, can get the best results for landuse evaluation. Every effort to improve knowledge of Pantanal is worthwhile, due to the fastchanges of vegetation cover caused by the flooding regime and also due to its conservation,particularly in Barão de Melgaço municipality, where there are two environmental reserves.

OBJECTIVES

The objectives of this study are: (1) to map the vegetation of Barão de Melgaçomunicipality during both rainy and dry seasons, applying two supervised classification methods(Parallelepiped and SVM) in MODIS images, as well as to evaluate the classifiers performance;(2) to compare the best result for each season with the map of phyto-ecologic regions ofthis municipality, and to qualitatively analyze the vegetation class.


MATERIALS AND METHODS

The municipality of Barão de Melgaço, located in the geographic micro-region UpperPantanal, has an area of approximately 11,175 km2. Figure 1 shows the localization of thestudy area inserted in the Brazilian Pantanal.

Figure 1 � Localization of the study area: Barão de Melgaço municipality,in Mato Grosso State

In this study the following MODIS data, flown onboard platform Terra, were used:product MOD13Q1, red (R), near infrared (NIR) and medium infrared (MIR) bands of tileh12v10, with 250 m spatial resolution. This product is a 16 day composition, and it is alreadyatmospherically corrected, i.e. it presents, as a quality parameter, the best pixels of the 16day image to compose the tile.

The images were acquired in accordance with the dry and rainy seasons of 2011. Forthe rainy season the image is Julian day 65th (composite from March 6th to 21st 2011) and forthe dry season the data taken was Julian day 209th (composite from July 28th to Aug. 12th

2011). For each period a color composite R(MIR) G(NIR) B(RED) as well as the selection ofthe study area were made.

The fraction images were generated for dry soil, vegetation, and water for the dryseason image, and for wet soil, vegetation, and water for the rainy season image. Theobjective was to enhance the targets with the contribution of each target contained in onlyone pixel, using the LSMM. The LSMM can be described according to equation (1) as follows:



Ri = sum (aij xj) + ei

Where:

Ri = medium spectral reflectance for the ith spectral band;

aij = spectral reflectance of the jth component in the pixel for the ith spectral band;

xj = proportion value of the jth component in the pixel;

ei = error for the ith spectral band;

j = 1,2�..n (n = number of components assumed for the problem)

i = 1,2�..n (n = number of spectral bands for the sensor system).

For the generation of fraction images, LSMM must select �pure pixels� of each target,also called �endmembers�, because they bring new information from the original spectralresponses.

For the dry season, pure pixels of dry soil, vegetation, and water were selected,while for the rainy season, pure pixels of wet soil, vegetation, and water were selected. Thisdifferentiation among pure soil pixels is due to the fact that this area presents large floodedsections during the rainy season and so the humid soil is more present and consequently itspectrally mixes with the vegetation, that is, the mapping target of this study. Figure 2shows the spectral curves of pure pixels selected, used at LSMM.

The following step was to apply two automatic supervised classification methods,namely Parallelepiped and SVM, and to extract the vegetation areas afterwards. Both classifierswere applied to each color composite image R(MIR) G(NIR) B(RED) and to the compositionsof fraction images R(wet soil) G(vegetation) B(water) and R (dry soil) G(vegetation) B(water),in order to verify if the utilization of fraction images improves the results of the classifications.

Figure 2 � Spectral curves of pure pixels of water, vegetation,wet soil and dry soil

The Parallelepiped method considers an area (square or rectangular) in the space ofattributes around the training set. According to Crósta (1993), each class has decisionborders, represented by the lateral sides of the parallelepiped. Therefore all pixels within thisarea are considered as belonging to the same class. The SVM method is based on the theoryof statistical learning (GONÇALVES et al., 2006). Its principle is the optimum separation ofclasses by a decision surface which maximizes the separation margin between such classes.This surface is also known as �ideal hyperplane� and the points which are close to the marginof the ideal hyperplane are called �support vectors� (BROWN et al., 2000). Even if SVM is abinary classifier in its simplest form, it can work as a multi-class classifier through thecombination of several binary classifiers SVM (PAN et al., 2012). In this study, the kernel RBF(Radial Basis Function) was used, which was also considered by Pan et al. (2012), Gonçalveset al. (2006) and Melgane & Bruzzone (2004).


In order to evaluate the accuracy of obtained maps, error matrices were generatedfor each classification. In an error matrix, the pattern of class attribution related to referencedata is described; i. e. inasmuch the situation considered is classified in conformity with thereference. This is a frequently used method for the accuracy evaluation of data originatedfrom remote sensing (FOODY, 2002). From the error matrix other important elements forsuch evaluation can be derived, such as overall accuracy, omission error, commission error,and Kappa coefficient (LU; WENG, 2007).

The references used were two thematic maps with generic vegetation classes; dry/wet soil and water, which were crossed with the classifications, using the post-classificationtool Confusion Matrix, Using Ground Truth ROIs, from software ENVI 4.5. These thematicmaps were generated from the visual interpretation of two Landsat-TM images with thecomposition R(5) G(4) and B(3). One image corresponds to the rainy season, obtained inApril 22nd 2011 and the other to the dry season, dated Aug. 12th 2011, close to the intervalof imaging dates from MODIS. In the literature there are several methodologies which considerthe above mentioned procedure, however, according to Mello et al. (2010), it is necessaryto consider the fact that the reference map does not truly express field reality, because theresult is a visual interpretation. So when the disagreement between the classification andthe reference is considered an �error�, it is wise to understand that the term is subjective.Among other studies, it is worth mentioning Araújo et al. (2011) and Shao & Lunetta (2012),who also followed a similar methodology for accuracy evaluation. The first authors used as areference a Landsat5-TM mosaic, while the second ones relied on a thematic map.

To reinforce the use of this methodology, this study was based on a discussion ofCongalton & Green (1999) where these authors discuss which information source should beconsidered as a reference, when one is working with data originated from satellite images.These authors inform that the reference data should be the one which considers the earlierstep, closer to the terrestrial reference, instead of the one which generated the classifiedmap. In this case, the Landsat-TM images are considered closer to the terrestrial conditionsthan the map generated from MODIS images.

Aiming at a qualitative analysis of the vegetation mapped, a comparison was madebetween the best vegetation classification obtained for each date and the map of phyto-ecological regions from the municipality. This map was elaborated as part of the activitiesfrom the sub-project �Survey and Mapping of Vegetation Cover Remnants from the PantanalBiome, in 2002, scale 1:250,000� which is inserted in the thematic �Survey of VegetationCover Remnants of Brazilian Biomes�, within the Project for Conservation and SustainableUse of the Brazilian Biologic Diversity, the PROBIO (BRASIL, 2006). The vegetation maskswere converted into vectors to cross with the vector data of the database from the map ofphyto-ecological regions.

RESULTS AND DISCUSSION

For each season four vegetation maps from Barão de Melgaço municipality weremade, resulting from the combination between the classifiers and the color composites ofimages, represented by the acronyms: PLP-FI referring to the Parallelepiped classifier appliedon fraction images; PLP-OB referring to the Parallelepiped classifier applied to compositionsR(MIR) G(NIR) B(RED), considered here as original bands (OB); SVM-FI referring to theclassifier SVM applied to compositions with fraction images and SVM-OB referring to theSVM classifier applied to the compositions R(MIR) G(NIR) B(RED).

The results of vegetation mapping with the image from the rainy season are presentedin figure 3.

Table 1 shows the statistical results derived from the error matrix for this mapping.



Figure 3 � Vegetation mapped during rainy season from combinations:(A) PLP-FI; (B) PLP-OB; (C) SVM-FI; (D) SVM-OB


The overall accuracy represents the total correct hits related to the total number ofpixels. The combination which presented the highest overall accuracy was PLP-OB, reaching89.5% correct hits. The lowest index of correct hits was obtained with SVM-OB, reaching83.9%.

Referring to the Kappa coefficient, which is a statistical technique for the evaluationof agreement or disagreement in two situations of interest, varying from 0 to 1 (CONGALTON,1991), the best result was obtained with SVM-FI, with Kappa value of 0.54. SVM-OB got thelowest Kappa: 0.43. All values of Kappa coefficient for the classifications of the rainy seasonare considered as of good quality, according to intervals proposed by Landis & Koch (1977).

The commission error occurs when other classes (in this case wet soil and water) areindentified by the classifier as the class of interest (vegetation). As for vegetation, which isthe object of the mapping, the commission error was low for all classifications, between1.5% and 2.0%, i.e. the other classes (wet soil and water) were rarely classified as vegetation.As a consequence, the user�s accuracy is the probability of a pixel classified in the map torepresent the same target in the terrestrial reference. For the vegetation class, the user�saccuracy varied between 98% and 98.5% for all classifications from the rainy season.

Finally table 1 also shows the omission error, which occurs when the class of interest(vegetation) is classified as the other classes (wet soil and water). Consequently theproducer�s accuracy is the probability that the pixel of interest is correctly classified. Forvegetation, the PBL-OB presented the lowest omission error (4.7%) and the highest producer�saccuracy (95.3%). SVM-OB came out with the highest omission error and the lowest producer�saccuracy, respectively 14.9% and 85.1%.

Antunes et al. (2011) showed that the rainfall regime in Pantanal determines analternation of the soil cover conditions. Comparing the temporal EVI profile with accumulatedprecipitation data, these authors observed that in areas without human occupation, thebiomass peaks coincide with the periods of higher rainfall, even when considering a timeinterval between the beginning of the rainy season and the response of vegetation.

Due to that, in those sections of dry soils during the dry season, the vegetationgrows after the beginning of rainfall, collaborating with the increase of the vegetationmapped area in the rainy season. Figure 4 illustrates vegetation mappings for the image ofAugust 2011, from the dry season. One observes a smaller area mapped with vegetationwhen compared with figure 3, of the rainy season. The accuracy values for figure 4 arepresented in table 2.

Table 1 � Accuracy evaluation for the map of March 2011,rainy season

* refers to class vegetation.



Figure 4 � Vegetation mapped during dry season from combinations:(A) PLP-FI; (B) PLP-OB; (C) SVM-FI; (D) SVM-OB


PLP-OB also obtained the best overall accuracy for the map of August, with 93.6% ofcorrect hits. On the other hand, PLP-FI obtained the lowest percentage of correct hits:83.2%.

Although the values of overall accuracy for the classifications from the dry seasonwere lower in comparison with those of the rainy season, the Kappa coefficient showed aninverse result. PLP-OB reached an excellent quality with Kappa of 0.87. The otherclassifications presented a very good quality with Kappa coefficients of 0.67 and 0.70. Thequality of classifications was evaluated according to the proposal from Landis & Koch (1977).

Analyzing the vegetation class, commission errors for image classifications from August2011 were higher compared with those of March 2011 classifications, i.e. the classifiersconfused dry soil and water with vegetation. So the user�s accuracy presented lower valuesof correct hits for this season. PLP-OB generated the lowest commission error (9.4%) andthe highest user�s accuracy (90.6%). On the other, hand SVM-FI presented 19.9% commissionerror and 80.1% user�s accuracy for the vegetation map during the dry season.

All other classifications, except for PLP-FI, presented a lower commission error and ahigher producer�s accuracy for the results from August, when compared to those fromMarch, i.e. few vegetation pixels were not classified correctly. The best result was reachedwith PLP-OB with 0.7% omission error and 99.3% producer�s accuracy.

Depending on the mapping objective, the overall accuracy can be sufficient to evaluateits quality. However, if some specific classes are important or are of higher interest thanothers, it is absolutely necessary to evaluate each class individually, taking into accountomission and commission errors and the user�s and producer�s accuracy. In this context,according to the results described, we conclude that the best method to map the vegetationfrom Barão de Melgaço municipality was PLP-OB, i. e. the Parallelepiped classifier applied onthe composition R(MIR) G(NIR) B(RED) for both rainy and dry seasons.

It was expected that the use of fraction images, derived from the Linear SpectralMixture Model (LSMM), would improve the classifiers performance, if compared to the use ofthe traditional original bands. According to Lu & Weng (2007) the presence of mixed pixels isone of the biggest problems which affects the image classification based on pixels. Classificationmethods which consider the sub-pixel have been developed to deliver a more appropriateland cover with higher accuracy, especially when using data with moderate to low spatialresolution, such as MODIS data.

It must be emphasized that, in this study, it was generally helpful to use the SVMclassifier with the fraction images, for the rainy season. A similar result was found byFerreira & Galo (2010) when they evaluated the influence of input data on the quality of

Table 2 � Accuracy evaluation for the map of August 2011,dry season

* refers to class vegetation.



Landsat-TM image classification by artificial neural networks. The analysis showed that useof fraction images with input data in the classification did not allow a significant increase onthe classifier performance, presenting a lower Kappa value in the results.

With these results in mind, it is interesting to highlight that the automatic satelliteimage classification does not depend only on the chosen classification method being adequatefor each objective. Lu & Weng (2007) report that this is a complex process which requiresthe consideration of several factors, and the most important are: the availability of highquality images and auxiliary data (e.g. relief and meteorological data), the design of theclassification procedure, and the abilities and experiences from the analyst. These authorspoint out that it is difficult to identify the best classifier for a specific study, due to the non-existence of both a guide for its selection and the availability of an adequate classificationalgorithm, which justifies the realization of comparative studies with different classifiers.

The best result for each season was qualitatively compared with the map of phyto-ecological regions of Barão de Melgaço (Figure 5). The main phyto-ecological regions in thismunicipality are Savanna Woodland (Sa), Savanna/Pioneer Formations (SP), Savanna/SeasonalForest (SN), Savanna Woodland + Tree/Shrub Savanna (Sa + Sg) and Seasonal Semi-Deciduous Alluvial Forest (Fa).

Figure 5B shows, among all vegetation classes, that only few sections of SavannaWoodland (Sa) were mapped during the dry season. According to the Technical Handbook ofBrazilian Vegetation (IBGE, 2012) this type of vegetation has a characteristic thinnanophanerophyte, a continuous hemi-crytophytic grass physiognomy. These dominant sinusesform a rachitic physiognomy in degraded terrains, which helps to understand the difficulty ofthe classifier, because there is a higher soil exposition during the dry season. Other vegetationtypes with poor mapping due to more bare soil were: Tree/Shrub Savanna (Sg), characterizedby grasses mixed with rachitic woody plants, the composed class Savanna Woodland +Tree/Shrub Savanna (Sa + Sg) and pasture areas which were formerly covered by Savanna(Ap.S).

The classes mapped with good performance are: Seasonal Forest/Pioneer Formations(NP), Savanna/Seasonal Forest (SN), Woodland (Sd) and Seasonal Semi-Deciduous AlluvialForest (Fa).

Figure 5C shows the vegetation mapped in the rainy season, divided in accordancewith the phyto-ecological regions. One observes that Savanna Woodland (Sa) had almostits entire area mapped. Savanna Woodland + Tree/Shrub Savanna (Sa + Sg) presented aslight improvement in its classification. This could be explained by the increase of vegetativevigor of these formations during the rainy season. Other vegetation types mapped with goodperformance during the rainy season were: Tree/Shrub Savanna (Sg), Woodland (Sd), SavannaWoodland + Tree/Shrub Savanna (Sd+Sa) and Seasonal-Semi-Deciduous Alluvial Forest (Fa).

One also observes that the classes Savanna/Seasonal Forest (SN), and PioneerFormations with fluvial or lacustrine influence (Pa) present a discontinuous mapping of itsvegetal formations when compared to the dry season, due to the formation of spatiallydistributed small lakes dispersed in this area during the rainy season.

As stated by Evans & Costa (2013), it is necessary to point out that the mappingderived from Project PROBIO was based on Landsat images, of improved spatial resolution,acquired in July and October, during the dry season as the authors did not consider thedynamics of flooding for the classification, not including, therefore, the component lake.

Although the objective of this study was not to individually map the phyto-ecologicalregions from Barão de Melgaço, this visual comparison can guide future studies, indicatingthe limitations and possibilities for each vegetal formation during the different seasons.


Figure 5 � Overlay of vegetation mapped with the phyto-ecological regions of Barãode Melgaço municipality. (A) Phyto-ecological regions; (B) Best classification for thedry season: PLP-OB; (C) Best classification for the rainy season: PLP-OBLegend: Ap.S: Cattle raising (pasture) / Savanna; Fa: Seasonal Semi-Deciduous Alluvial Forest;NP: Seasonal Forest / Pioneer formations; Pa: Pioneer formation with fluvial and/or lacustrineinfluence; SN: Savanna / Seasonal forest; SP: Savanna / Pioneer Formations; Sa: SavannaWoodland; Sa + Sd: Tree/Bush Savanna + Woodland; Sa + Sg: Woodland + Tree/Shrub Savanna;Sd: Woodland; Sd + Sa: Woodland + Woodland Savanna; Sd+Sg: Woodland + Tree/Shrub Savanna;Sg: Tree/Shrub Savanna; Sg+Sa: Tree/Shrub Savanna + Savanna Woodland; Vs.S: NaturalSecondary Succession / Savanna.



CONCLUSIONS AND SUGGESTIONS

Mapping vegetation of Barão de Melgaço municipality, during both rainy and dryseasons, presented quality indices between good and excellent. In spite of that, the use offraction images derived from the Linear Spectral Mixture Model as input data for theclassifications did not present the result expected to improve the performance of the classifiers,except for the SVM applied on the image fraction composition for the rainy season, whichobtained the best Kappa (0.54) for such season.

Considering the vegetation as the target of interest in the mapping, the best resultswere obtained by the Parallelepiped classifier applied on the original spectral bands of sensorMODIS for both seasons. Although during the rainy season the vegetation is more vigorous,mapping during the dry season presented a higher level of correct hits and a lower level oferrors due to the improvement of the spectral discrimination among the targets.

The qualitative comparison between the classifications and the phyto-ecologicalregions from the municipality showed that those smaller classes mapped during the dryseason were the Woodland Savanna and the Tree/Bush Savanna, considering their physicalcharacteristics. During the rainy season, due to the increase of the vegetative vigor, suchclasses could be identified.

REFERENCES

ADAMI, M.; FREITAS, R. M.; PADOVANI, C. R.; SHIMABUKURO, Y. E.; MOREIRA, M. A. Estudoda dinâmica espaço-temporal do bioma Pantanal por meio de imagens MODIS. PesquisaAgropecuária Brasileira (Online), Brasília, v. 43, n. 10, p. 1371 - 1378, 2008. Available at:< http://www.scielo.br/pdf/pab/v43n10/16.pdf> Access in Dec. 9th 2012.

ADAMOLI, J. Zoneamento ecológico do Pantanal baseado no regime de inundações. In:ENCONTRO SOBRE SENSORIAMENTO REMOTO APLICADO A ESTUDOS NO PANTANAL, 1.,Corumbá, 1995. Anais... São José dos Campos: INPE, 2005. p. 15-17.

ANTUNES, J. F. G. ; ESQUERDO, J. C. D. M. ; LAMPARELLI, R. A. C. Temporal dynamicmonitoring of four vegetation cover of Pantanal using the Wavelet Transform applied to atime-series of EVI/MODIS data. Geografia, Rio Claro, v. 36, p. 173 - 185, 2011.

ARAÚJO, G. K. D.; ROCHA, J. V.; LAMPARELLI, R. A. C.; ROCHA, A. M. Mapping of summercrops in the State of Paraná, Brazil, through the 10-day spot vegetation NDVI composites.Engenharia Agrícola, Jaboticabal, v. 31, n. 4, p. 760-770, 2011.

BRASIL. Ministério do Meio Ambiente. Cobertura Vegetal dos Biomas Brasileiros (anobase 2002). Folhas SE.21-X-A Poconé e SE.21-X-C Ilha Camargo. Brasília, 2006. Availableat <http://www.macroprograma1.cnptia.embrapa.br/projeto/probiopantanal/downloads-1>Access in June 29th 2012.

BROWN, M.; LEWIS, H. G.; GUNN, S. R. Linear Spectral Mixture Models and Support VectorMachines for Remote Sensing. IEEE Transactions on Geoscience and Remote Sensing,New York, v. 38, n. 5, p. 2346 � 2360, 2000.

CONGALTON, R.G. A review of assessing the accuracy classifications of remotely senseddata. Remote Sensing of Environment, New York, v. 37, p. 35 - 46, 1991.

CONGALTON, R.G.; GREEN, K. Assessing the Accuracy of Remotely Sensed Data: Principlesand Practices. (Mapping Sciences Series). New York: CRC Press, 1999. Cap. 4, p. 27-41.


CRÓSTA, A. P. Processamento Digital de Imagens de Sensoriamento Remoto. Ed. rev.Campinas, SP: IG/UNICAMP, 1993. 170p.

DEVI, M. R.; BABOO, S. S. Land Use and Land Cover Classification using RGB&L BasedSupervised Classification Algorithm. International Journal of Computer Science &Engineering Technology, v. 2, n. 10, p. 167 � 180, 2011.

EVANS, T. L.; COSTA, M. Land cover classification of the Lower Nhecolândia sub-region ofthe Brazilian Pantanal Wetlands using ALOS/PALSAR, RADARSAT-2 and ENVISAT/ASAR imagery.Remote Sensing of Environment, New York, v. 128, p. 118 � 137, 2013.

FERREIRA, M. S.; GALO, M. L. B. T. Influência dos dados de entrada na classificação a partirde rede neural artificial. In: SIMPÓSIO BRASILEIRO DE CIÊNCIAS GEODÉSICAS E TECNOLOGIASDA GEOINFORMAÇÃO, 3., Recife, 2010. Anais... Recife: UFPE, 2010. p. 1 - 7.

FOODY, G. M. Status of land cover classification accuracy assessment. Remote Sensing ofEnvironment, New York, v. 80, p. 185 - 201, 2002.

GONÇALVES, P.; CARRÃO, H.; PINHEIRO, A.; CAETANO, M. Land cover classification withSupport Vector Machine applied to MODIS imagery. In: EARSeL SYMPOSIUM, 25., Porto,Portugal, 2005. Proceedings� Rotterdam: Millpress, 2006. p. 517 � 525. Available at: <http://www.earsel.org/symposia/2005-symposium-Porto/index.htm> Access in Dec. 9th 2012.

IBGE. INSTITUTO BRASILEIRO DE GEOGRAFIA E ESTATÍSTICA. Manual Técnico da VegetaçãoBrasileira. 2ª ed. Rio de Janeiro: IBGE, 2012. 271p. Available at: <ftp://geoftp.ibge.gov.br/documentos/recursos_naturais/manuais_tecnicos/manual_tecnico_vegetacao_brasileira.pdf>Access in Dec. 28th 2012.

JENSEN, J. R. Sensoriamento Remoto do Ambiente: uma perspectiva em recursos terrestres.Translation of 2nd edition by: EPHIPHANIO, J. C.; FORMAGGIO, A. R; SANTOS, A. R.; RUDORFF,B. F. T.; ALMEIDA, C. M.; GALVÃO, L. S. São José dos Campos: Parêntese, 2009. 672p.

JUNK, W. J.; CUNHA, C. N.; WANTZEN, K. M.; PETERMANN, P.; STRÜSSMANN, C.; MARQUES,M. I.; ADIS, J. Biodiversity and its conservation in the Pantanal of Mato Grosso, Brazil.Aquatic Sciences, Basel, v. 68, p. 278 - 309, 2006.

LANDIS, J. R.; KOCH, G. G. The measurement of observer agreement for categorical data.Biometrics, Washington, v. 33, n. 1, p. 159 - 174, 1977.

LATORRE, M. L.; SHIMABUKURO, Y. E.; ANDERSON, L. O. Produtos para Ecossistemas Terrestres- MODLAND. In: RUDORFF, B. F. T.; SHIMABUKURO, Y. E.; CEBALLOS, J. C. (Org.) O SensorMODIS e suas Aplicações Ambientais no Brasil. São José dos Campos: Parêntese, 2007.Cap. 2, p. 23 � 35.

LU, D.; WENG, Q. A survey of image classification methods and techniques for improvingclassification performance. International Journal of Remote Sensing, Basingstoke, v. 28,n. 5, p. 823 � 870, 2007.

MELGANI, F.; BRUZZONE, L. Classification of Hyperspectral Remote Sensing Images WithSupport Vector Machines. IEEE Transactions on Geoscience and Remote Sensing, NewYork, v. 42, n. 8, p. 1778 � 1790, 2004.

MELLO, M.P.; RUDORFF, B.F.T.; VIEIRA, C.A.O.; AGUIAR, D.A. de. Classificação automáticada colheita da canadeaçúcar utilizando modelo linear de mistura espectral. Revista Brasileirade Cartografia, Rio de Janeiro, n.62/2, p.181188, 2010.

PAN, Y.; HU, T.; ZHU, X., ZHANG, J., WANG, X. Mapping Cropland Distributions Using a Hardand Soft Classification Model. IEEE Transactions on Geoscience and Remote Sensing,New York, v. 50, n. 11, p. 4301 � 4312, 2012.



SENTHILNATH, J.; BAJPAI, S.; OMKAR, S. N.; DIWAKAR, P. G.; MANI, V. An approach tomulti-temporal MODIS image analysis using image classification and segmentation. Advancesin Space Research, Oxford, v. 50, p. 1274 � 1287, 2012.

SHAO, Y.; LUNETTA, R. S. Comparison of support vector machine, neural network, and CARTalgorithms for the land-cover classification using limited training data points. ISPRS Journalof Photogrammetry and Remote Sensing, Amsterdam, v. 70, p. 78 � 87, 2012.

SHIMABUKURO, Y. E.; NOVO, E. M.; PONZONI, F. J. Índice de Vegetação e Modelo Linear deMistura Espectral no Monitoramento da Região do Pantanal. Pesquisa AgropecuáriaBrasileira, Brasília, v. 33, número especial, p. 1729 - 1737, 1998.

SHIMABUKURO, Y. E.; SMITH. J. A. The least-squares mixing models to generate fractionimages derived from remote sensing multispectral data. IEEE Transactions on Geoscienceand Remote Sensing, New York, v. 29, n. l, p. 16 - 20, 1991.

SILVA, J. S. V.; ABDON, M. M. Delimitação do Pantanal Brasileiro e suas Sub-regiões. PesquisaAgropecuária Brasileira, Brasília, v. 33, número especial, p. 1703 - 1711, 1998.

SILVA, G. B. S. da, FORMAGGIO, A. R.; SHIMABUKURO, Y. E.; ADAMI, M.; SANO, E. E.Discriminação da cobertura vegetal do Cerrado Matogrossense por meio de imagens MODIS.Pesquisa Agropecuária Brasileira, Brasília, v. 45, n. 2, p. 186 - 194, 2010.

VIANA, D. R.; ALVALÁ, R. C. dos S. Vegetation Index Performance for the Pantanal RegionDuring Both Dry and Rainy Seasons. Geografia, Rio Claro, v. 36, p. 143 - 158, 2011.

ZHAN, X.; SOHLBERG, R. A.; TOWNSHEND, J. R. G.; DIMICELI, C.; CARROLL, M. L.; EASTMAN,J. C.; HANSEN, M. C.; DEFRIES, R. S. Detection of land cover changes using MODIS 250 mdata. Remote Sensing of Environment, New York, v. 83, p. 336 � 350, 2002.

supervised classification applied to vegetation mapping in the barˆo de … · 2014. 5. 8. ·...

Documents